Process teardown hang via circular workqueue wait

Improper locking in the Linux kernel amdkfd SVM code allows local users to hang process teardown by exiting a GPU process. The kfd_process_notifier_release handler synchronously cancels SVM restore work while that worker waits for deferred work whose final mmput re-enters the same release path, forming a circular wait. Reachability requires low-privilege local access to /dev/kfd and use of SVM or unified memory; the deadlock retains process and GPU resources and consumes workqueue workers.